Continuous and real-time learning is a difficult problem in robotics. To le
arn efficiently, it is important to recognize the current situation and lea
rn appropriately for that context. To be effective, this requires the integ
ration of a large number of sensorimotor and cognitive signals. So far, few
principles on how to perform this integration have been proposed. Another
limitation is the difficulty to include the complete contextual information
to avoid destructive interference while learning different tasks.
We suggest that a vertebrate brain structure important for sensorimotor coo
rdination, the cerebellum, may provide answers to these difficult problems.
We investigate how learning in the input layer of the cerebellum may succe
ssfully encode contextual knowledge in a representation useful for coordina
tion and life-long learning. We propose that a sparsely-distributed and sta
tistically-independent representation provides a valid criterion for the se
lf-organizing classification and integration of context signals. A biologic
ally motivated unsupervised learning algorithm that approximate such a repr
esentation is derived from maximum likelihood. This representation is benef
icial for learning in the cerebellum by simplifying the credit assignment p
roblem between what must be learned and the relevant signals in the current
context for learning it. Due to its statistical independence, this represe
ntation is also beneficial for life-long learning} by reducing the destruct
ive interference across tasks, while retaining the ability to generalize. T
he benefits of the learning algorithm are investigated in a spiking model t
hat learns to generate predictive smooth pursuit eye movements to follow ta
rget trajectories.